Handy AI is a reader-supported publication. To receive new posts and support our work, consider becoming a free or paid subscriber.
what to know for now
🖼️ ChatGPT’s new image generation lands in API. The upgraded image engine behind ChatGPT now lives in the API, freeing devs to jam generative art straight into their pipelines. Rate limits rose; watermarking stays. Read more
💻 Microsoft 365 Copilot April wave. Copilot Pages can export to Word, Researcher and Analyst agents join Planner and SharePoint, and a Win + C shortcut launches context-aware chat that reads on-screen pixels for summarization and rewrite. IT gains Copilot Control System APIs for policy and usage telemetry. Rollout starts in May across enterprise tenants. Read more
🛠️ xAI launches Grok Studio. Grok Studio lets users draft docs, run Python, and edit Google Drive files while Grok chats on the left. A sandbox container streams stdout to the panel, so prototypes compile without context-switching. Free-tier access aims to grow share against ChatGPT’s paid canvas. Read more
🎨 Adobe unveils Firefly mobile app for on-device generative imaging. The iOS and Android beta packs the latest Firefly model with 4-step diffusion, editable vector layers and on-chip semantic masking. Assets sync to Creative Cloud for round-trip into Photoshop and Illustrator. Read more
Jake's Take: This study gathered consistent tags from professional annotators on film-school rules, then pitted Structure-from-Motion stacks and vision-language models against each other. A fine-tuned Qwen 2.5-VL hit the same score as the strongest geometry tracker, proving text supervision can teach motion sense.
CameraBench (the benchmark created in the study) exposes failures in both SfM and vision-language models. SfM falls apart once the scene starts moving, while vision-language models miss precise angles. This implies that both camps should always fuse pose math and language reasoning, inject camera moves into captions, and let video generators plan real shots.
what to know for later
🧘 Anthropic launches research track on model welfare. Anthropic opened a research track on model welfare after alignment teams raised questions about agency, preferences and distress in language models. The project will build probes, protocols and interventions and feed results back into Claude training. A report led by David Chalmers says some future systems could merit moral consideration. Read more
🤖 OpenAI prototypes an ‘open’ reasoning model. The experimental architecture keeps a slim local model that routes hard prompts to stronger GPT-4-class endpoints, yielding faster first-token latency on device while preserving access to full-scale accuracy when needed. OpenAI says the handoff uses a secure RPC layer with streaming, so tokens flow back into the local context window in real time. Developers will test paid tiers later this quarter. Read more
📰 The Washington Post inks content-licensing deal with OpenAI. ChatGPT will surface attributed Post excerpts, headlines and links, broadening OpenAI’s push to fold premium journalism into its RAG pipeline. The pact covers current and archival stories and mirrors earlier agreements with Axel-Springer and Le Monde, showing publishers tolerate LLMs. Both firms promise click-through telemetry and revenue sharing. Read more
🎓 White House mandates K-12 AI curriculum. An executive order titled “Advancing Artificial Intelligence Education for American Youth” tasks the Departments of Education and Commerce with open-sourcing grade-level modules on model literacy and ethics. States that align standards gain priority in NSF research grants. Vendors eye a textbook refresh cycle and Chromebook fleet upgrades. Read more
🍏 Apple breaks up its AI org. Internal memos leaked to Bloomberg cite duplicated research and slow shipping cadence. The shuffle places most LLM work under John Giannandrea, while ex-robotics staff join the Vision Products Group. Read more
